Home Knowledge Base AI/ML Glossary

AI/ML Glossary

Core LLM Concepts

A-F

TermDefinition
AlignmentTraining AI to be helpful, harmless, and honest
AttentionMechanism for tokens to consider other tokens
BPEByte Pair Encoding, a tokenization algorithm
Chain-of-ThoughtPrompting technique for step-by-step reasoning
Context WindowMaximum tokens an LLM can process at once
DecoderTransformer block that generates output tokens
DPODirect Preference Optimization, simpler RLHF alternative
EmbeddingDense vector representation of text
EncoderTransformer block that processes input
Fine-tuningAdapting a pretrained model to new data
Few-shotProviding examples in the prompt

G-L

TermDefinition
Ground TruthCorrect labels for training or evaluation
HallucinationLLM generating plausible but false information
InferenceRunning a trained model to get predictions
JailbreakCircumventing LLM safety measures
KV CacheStored key-value pairs for efficient generation
LoRALow-Rank Adaptation, parameter-efficient fine-tuning
LLMLarge Language Model
LossMeasure of prediction error during training

M-R

TermDefinition
MoEMixture of Experts architecture
MultimodalProcessing multiple data types (text, image, audio)
PerplexityExponential of cross-entropy, measures uncertainty
Prefix CachingReusing cached KV for common prefixes
PromptInput text given to an LLM
QuantizationReducing numeric precision (FP16 → INT4)
RAGRetrieval-Augmented Generation
RLHFReinforcement Learning from Human Feedback
RoPERotary Position Embedding

S-Z

TermDefinition
SFTSupervised Fine-Tuning on instruction data
Speculative DecodingUsing draft model to accelerate generation
System PromptInstructions defining AI behavior
TemperatureControls randomness in generation
TokenSubword unit processed by LLM
Top-pNucleus sampling parameter
TransformerNeural network architecture with attention
TTFTTime to First Token
VLMVision-Language Model
Zero-shotPrompting without examples

Infrastructure Terms

TermDefinition
CUDANVIDIA's GPU computing platform
Flash AttentionMemory-efficient attention algorithm
HBMHigh Bandwidth Memory (GPU memory)
NVLinkHigh-speed GPU interconnect
TensorRTNVIDIA inference optimization library
vLLMHigh-throughput LLM serving engine
GGUFFile format for quantized models

Metrics

TermDefinition
BLEUMachine translation quality metric
F1Harmonic mean of precision and recall
Pass@kCode generation success probability
TPOTTime Per Output Token
WERWord Error Rate for speech recognition
glossarydefine termsjargonconcept

Explore 500+ Semiconductor & AI Topics

From EUV lithography to CUDA optimization — search the full knowledge base or chat with our AI assistant.